🔧 (CI): Use cibuildwheel for wheel building

[Isotr0py]

Related issue: #74

[jfolz]

Re: qemu.

Since it's so terribly slow, I implemented separate jobs for different architectures (and PyPy) for one of my projects, see build_wheels.yml. The matrix is terrible, but it illustrates the concept :)

[Isotr0py]

Not sure if the regression performance related to my device, the performance of built wheel from CI with vendored looks good on another device:

/content/pillow-jpegxl-plugin
Quality: 98

Effort 1
  plugin    [size: 258714, duration: 0.408]
  reference [size: 258716, duration: 0.346]

Effort 2
  plugin    [size: 258714, duration: 0.327]
  reference [size: 258716, duration: 0.347]

Effort 3
  plugin    [size: 240657, duration: 0.482]
  reference [size: 240659, duration: 0.355]

Effort 4
  plugin    [size: 242940, duration: 0.585]
  reference [size: 242942, duration: 0.720]

Effort 5
  plugin    [size: 342423, duration: 0.943]
  reference [size: 342425, duration: 0.650]

Effort 6
  plugin    [size: 341787, duration: 0.764]
  reference [size: 341789, duration: 0.900]

Effort 7
  plugin    [size: 340736, duration: 0.825]
  reference [size: 340738, duration: 0.985]

Effort 8
  plugin    [size: 349518, duration: 4.144]
  reference [size: 349520, duration: 3.226]

Effort 9
  plugin    [size: 352912, duration: 4.956]
  reference [size: 352914, duration: 5.897]

Therefore, I'm going to merge this PR for Linux building CI first, and migrate Windows and MacOS CI in a seperate PR. :)
I will make a new release for python3.13 after supporting I;16 and F modes as well.

@jfolz @rphlo Thanks for your help very much!